Scholarlink

<p>Accurate segmentation of cutaneous manifestations in Systemic Lupus Erythematosus (SLE) is important for objective disease activity assessment in clinical practice. However, achieving both high segmentation accuracy and practical inference efficiency remains challenging, particularly in real-time or near real-time clinical workflows. In this work, we propose BD-ResTransUNet, a boundary-driven dual-stream framework designed for practical real-time medical image segmentation, achieving a favorable balance between accuracy and computational cost. The proposed method explicitly separates boundary-aware representation learning from semantic feature extraction and integrates them through a geometric-adaptive fusion mechanism. Specifically, an explicit boundary stream introduces structural priors to enhance boundary localization, while geometric-adaptive hybrid deformable attention modules enable flexible modeling of non-rigid lesion structures with reduced computational overhead. In addition, a multi-scale fusion module is employed to align complementary features across different resolutions. Experiments conducted on the SLE dataset and three additional public benchmarks (ISIC2016, SLHM, and MFSD) demonstrate the effectiveness of this integrated paradigm. BD-ResTransUNet achieves competitive segmentation accuracy with a Dice score of 0.872 ± 0.034 and a reduced HD95 of 29.0 ± 1.2 on the SLE dataset. The framework runs at 20.4 FPS (49.1&#xa0;ms per image) with 75.5&#xa0;M parameters on GPU hardware, demonstrating a practical balance between accuracy and inference efficiency. These results suggest that the proposed framework provides an effective and transferable solution for medical image segmentation tasks requiring both structural precision and practical real-time computational performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

BD-ResTransUNet: an efficient boundary-driven framework for real-time segmentation of SLE lesions

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Kun Zhu,
  • Nur Hana Samsudin

摘要

Accurate segmentation of cutaneous manifestations in Systemic Lupus Erythematosus (SLE) is important for objective disease activity assessment in clinical practice. However, achieving both high segmentation accuracy and practical inference efficiency remains challenging, particularly in real-time or near real-time clinical workflows. In this work, we propose BD-ResTransUNet, a boundary-driven dual-stream framework designed for practical real-time medical image segmentation, achieving a favorable balance between accuracy and computational cost. The proposed method explicitly separates boundary-aware representation learning from semantic feature extraction and integrates them through a geometric-adaptive fusion mechanism. Specifically, an explicit boundary stream introduces structural priors to enhance boundary localization, while geometric-adaptive hybrid deformable attention modules enable flexible modeling of non-rigid lesion structures with reduced computational overhead. In addition, a multi-scale fusion module is employed to align complementary features across different resolutions. Experiments conducted on the SLE dataset and three additional public benchmarks (ISIC2016, SLHM, and MFSD) demonstrate the effectiveness of this integrated paradigm. BD-ResTransUNet achieves competitive segmentation accuracy with a Dice score of 0.872 ± 0.034 and a reduced HD95 of 29.0 ± 1.2 on the SLE dataset. The framework runs at 20.4 FPS (49.1 ms per image) with 75.5 M parameters on GPU hardware, demonstrating a practical balance between accuracy and inference efficiency. These results suggest that the proposed framework provides an effective and transferable solution for medical image segmentation tasks requiring both structural precision and practical real-time computational performance.